Skip to main content
Music Maker AI
Open model · 2025

DiffRhythm AI song generator

Give it lyrics with a time on every line and a style to copy. Try the makers’ demo right here.

DiffRhythm comes from ASLP-lab and sings in Chinese or English. Its demo wants each lyric line written as [mm:ss.xx] and makes between 95 and 285 seconds of music, styled from a short audio clip or from a text prompt.

ASLP-lab hosts this demo on Hugging Face, where free machines are shared and a run may wait or fail. Nothing here touches your credits, and it isn’t part of our site.

Open on Hugging Face
  • 95 to 285 seconds
  • A time on every lyric line
  • Chinese and English

You decide when each line lands

Most song makers guess the timing. Here you set it, so a badly timed line is yours to fix. The picture is “Lisbon in Our Eyes”, from our own generator, as an example of a gentle acoustic song to pace against.

A spectrogram of “Lisbon in Our Eyes”, with soft stripes of colour low down where a guitar and voice sit.

Write the time first

A valid line looks like [00:10.00] and then the words. Lines without a time won’t follow the format.

Stay under the cap

The last time in your lyrics can’t be later than 04:45.00.

No lyrics yet?

The demo has a second tab for writing them, so you can start from a theme.

Two ways to choose the sound

The demo gives you an audio prompt or a text prompt. These two songs from our generator show the kind of contrast to ask for.

An orange waveform of “Gone With the Rain”, with quiet verses opening up into louder sections.

Bring a clip

Upload at least a second of audio; anything over 10 seconds is cut to a random 10. Pick those seconds with care, since short clips can make a song fall apart.

A spectrogram of “Kitchen Floor”, with dense warm colour at the bottom and thin bright lines above.

Or describe it

Switch to the text prompt for something like “indie folk ballad, acoustic guitar with harmonica”. Only use a clip you have the right to use.

DiffRhythm at a glance

As shown in the demo’s own guide and on its model page.

Song length
95 to 285 seconds (1:35 to 4:45)
Languages
Chinese and English; more promised
Style from
A clip of up to 10 seconds, or text
Modes
Quality first or speed first
Slow to load?
The guide suggests picking MP3 in Advanced Settings
Paid use
Unclear: the page names Apache 2.0 and also Stable Audio’s community terms

Singing from our own generator

Not DiffRhythm: these came from our song maker, with no times typed in.

Cover of “Lisbon in Our Eyes”: a yellow tram on a sunny Lisbon hill

Lisbon in Our Eyes

2:29

Sweet acoustic love song · male vocals

  • Acoustic
  • Male vocals
  • 2:29

Made from: “sweet acoustic love song, fingerpicked guitar, warm male vocal, anniversary”

Cover of “Kitchen Floor”: a pot of pasta and a radio in a warm kitchen at night

Kitchen Floor

2:30

Smooth R&B love song · male vocals

  • R&B
  • Male vocals
  • 2:30

Made from: “smooth R&B love song, silky vocals, warm keys, slow groove, romantic”

DiffRhythm questions

How do I format lyrics for DiffRhythm?

Put the start time in square brackets before each line, as in [00:13.20] and then your words. An .lrc lyric file already looks like this, so you can paste one straight in.

Can DiffRhythm make a song shorter than 95 seconds?

Not in this demo: 95 seconds is the bottom of its slider. For something like a 30-second jingle, a different model or tool is the better choice.

Is DiffRhythm safe for paid projects?

We can’t say. Its page names two sets of terms that don’t say the same thing, so ask its makers before you sell work made with it.

Not covered here.Ask us something else

Rather not type timestamps?

Paste plain lyrics into our maker and it works out where each line falls.

Make a song

Two free songs, no account needed.